Deep neural networks are increasingly used in safety-critical systems, but a small number of bit flips in memory can cause a trained model's accuracy to collapse. Attackers can induce such flips deliberately, and existing defensive approaches tend to be expensive: they add significant hardware overhead, rely on retraining, or fail against carefully targeted flips.
A team from the University of Florida, the University of Texas at Dallas, and Rensselaer Polytechnic Institute has proposed BARE-AI, a runtime framework described in an October 2026 arXiv paper. Rather than modifying the model or adding dedicated circuits, BARE-AI takes advantage of built-in performance monitors found in AI hardware. During inference, it watches for anomalies that indicate bit flips, then localizes the affected area and mitigates the fault on the fly.
The paper, 'BARE-AI: Bit-Flip Attack Resilience in AI Hardware through Built-in Performance Monitors,' was authored by Habibur Rahaman, Swastik Bhattacharya, Sanjay Das, Kanad Basu, and Swarup Bhunia. The abstract reports that the framework detects, localizes, and mitigates attacks during inference; because the source provides only the abstract, specific performance numbers and comparisons with prior work are not available here.